1. Field of the Invention
This invention generally relates to packet communications switching and, more particularly, to a cut-through packet routing system and method that has been optimized to reduce latency through the switch fabric.
2. Description of the Related Art
There is industry demand for integrated circuits (ICs) switching systems that can be easily scaled for Network Access, the Network Edge, or a Core Switch Router, at the high end. SCSA (Signal Computing System Architecture) defines a switch fabric to be the facility for connecting any two (or more) transmitting or receiving Service Providers.
FIG. 13 is a schematic block diagram of a general switch fabric system (prior art). Traffic enters the system on the ingress side (the left side of the figure). Traffic exits the system on the egress side (the right side of the figure). The port card ingress and egress structures are sometimes subdivided into channels and subchannels. The packet headers specify a required destination port card, channel and subchannel. The system must synchronously route each ingress packet to the correct egress destination.
Packets are converted into frames by ingress traffic managers (iTMs). A frame is a logical unit of data, which is often a small piece of a much larger data set such as a file or image. The iTMs feed the ingress side of the fabric. The switch fabric might convert the frame format to a “native” format, and then on egress, convert the data back into the TM frame format before sending the data to the egress traffic managers (eTMs). If the frames are fixed size (for example: 53 bytes, 64 bytes, or 80 bytes), the frames are often called cells.
Protocol Specific Vs. Protocol Agnostic
A switch fabric can be protocol specific or protocol agnostic. An example of a protocol specific switch fabric would be a system designed and optimized specifically for asynchronous transfer mode (ATM) traffic. Another example would be a switch that handles only TCP/IP traffic. The obvious disadvantage of a switch fabric hardwired to handle a specific protocol is lack of flexibility. Service providers want to sell services to a variety of customers in various industries. Protocols vary from industry to industry.
Even within one protocol type, there can be protocol upgrades. For example, TCP/IP, the machine language of Internet routers, now runs primarily “IPv4.” This protocol, successful as it has been, has a lack of available addresses, poor security features, and no “quality of service” (QoS) provisions. The next generation Internet protocol is “IPv6.” It provides solutions to these limitations.
A protocol agnostic switch fabric works equally well with all protocol types, however, the traffic manager must be responsible for recognizing specific protocol types. The disadvantage of a protocol agnostic switch fabric is that it may be more complicated, and perhaps slower than a switch fabric dedicated to a particular protocol.
Packet Striping Vs. Single Link Per Packet
The simplest way for a traffic manager to transmit a packet into a switch fabric is to transmit the packet serially along one line. Striping is a way of achieving higher bandwidth by transmitting a single packet across multiple ingress/egress lines. For example, a TM can transmit a packet into a switch fabric eight times as fast if the packet is sliced into eight pieces (stripes) by the TM, and conveyed into the fabric along 8 parallel lines simultaneously. The fabric captures the packet in memory, routes it to the required egress destination, and slices the packet into 8 parallel lines before transmitting the packet to the egress TM.
The upside to packet striping is the potential for lower latency. There are several negative aspects of packet striping:
if one of the links is damaged (1 of 8 in the example above), the entire channel is out of service, degrading fault tolerance; and,
the interface between the TM and switch fabric is more complicated. Circuitry must be used to slice the packet into stripes and reassemble it into packets.
Single-Cell Packet Vs. Multi-Cell Packets
Many switch fabrics now deployed and passing revenue traffic, especially in wide area networks (WANs), use asynchronous transfer mode (ATM) packets. ATM packets are single-cell packets, with a fixed cell size of 53 bytes, 48 bytes of which is the information payload. The ATM specification evolved in the 1980s and early 1990s. It was then believed that variable length (multi-cell) packets would be too difficult to implement at rates needed for wire-speed traffic. The single-cell solution was a compromise that would work for voice data, video data, multimedia data, email data, file data, etc. With a fixed frame size, switch designs are simplified and ultimately faster.
However, files are better sent in large frames. Voice switching performs better with small bursts of data, corresponding to analog-to-digital converter time slots. Large frames and concomitant switching latencies can render a switch useless for most 2-way voice applications. Voice communications require low latency (time delay). Since ATM had to work for all network data, the small payload (voice) requirements prevailed at the expense of the large frame applications.
For large frames or packets it is much more efficient to establish a path from an ingress port card to the required egress destination, and then leave this path undisturbed until the packet is completed. With single cell packets, the ingress port card must bid for, accept grants, and then schedule each cell of a multi-cell packet.
There are also Frame Relay switch fabrics. Frame Relay is a single-cell protocol, albeit with frame size ranging from 6 to 4096 bytes. Such switch fabrics have been deployed and passing revenue traffic since the early 1990s. These switch fabrics have generally not been used for voice data because of the large latency. Only highly compressed voice traffic works well over frame relay. Voice-over-frame relay was added as an afterthought by network engineers. Frame relay excels at data communications such as local area network internetworking (LAN-to-LAN). Such communications are very high speed and bursty, with non-critical latency constraints.
Cut-Through Vs. Store-and-Forward
The conventional method of switch fabric packet routing is called Store-and-Forward. In this method, the switch fabric accepts an input packet and buffers the packet on the ingress side of the fabric, making sure the packet was received intact, knowing the exact number of cells in the packet. The problem with the store-and-forward method is the added latency of buffering the packet. In Cut-through Packet Routing (CPR), a switch fabric is able to send the incoming packet cells to the correct egress port as soon as the destination address is known.
Memory-Based Vs. Arbitrated Crossbar
A switch fabric can use memory-based crossbars or arbitrated crossbars. A memory-based crossbar is sometimes called a “shared memory switch.” Ingress packets flow from the port cards into a huge memory bank, which serve as the switch. From the memory bank, the destination address is determined from egress port destination information in the cell headers. The problem with these switch fabrics is that they become prohibitively expensive and large from all the required high-speed memory. Such fabrics cannot be used to reach terabit total switching speeds with technology available today.
It would be advantageous if a switch fabric backplane flow management could be conducted without interrupting the backplane traffic flow.